Files
rust-discovery/docs/apt-ambient-reboot.md
T
2026-11-13 20:12:30 +01:00

168 lines
9.6 KiB
Markdown

# APT host reboot evidence and the configured historical attempt
Actual result on 2026-11-13: the Cloud Linux VM's historical configure attempt
`33775478-15b0-46ec-b059-44f500dbb0b6` is now **failed** with
`APT_RECOVERY_CONFIGURED_ACCEPTANCE_REBOOT_FLAG_VERIFIED`. The package effect
completed, but this is a failed acceptance, not an accepted installation.
Normal signed RootSync closed the same attempt; no new job or package command
was used for this classification. The notebook remains development-only and
the server remains `discovery.ease-technologies.de`.
## Why the earlier result was unknown
The separately WebUI-approved configure-only job configured fixture `1.0` and
returned APT exit 0 with the expected hook armed. Its private sandbox reported
no reboot marker. The parent observed the host's reboot-required marker, so
the previous provider returned `FIXTURE_UNEXPECTED_REBOOT_REQUIRED` and kept
the attempt unknown. That provider had not measured the host flag before the
effect. A private `/run` observation cannot establish the host's earlier state.
The marker's timestamp does not replace that missing measurement.
The original failed unpacked attempt
`31ef5e5e-5dbd-4428-a9a1-cbd15de2f5a0`, its partial proof
`dc3bf3870ea77c12f9291bac712c53ae72194eaec980a8646155a7fb5c0dbe21`,
both bootstrap tombstones and the configure attempt marker remain intact.
Neither an install acceptance nor a recovery acceptance was created for 337.
## New provider contract
Spec 1.13 permits a fully verified package result with an already pending host
reboot. For future mutating jobs, the parent now seals the host flag and boot ID
before entering the sandbox. Native prechecks schema 2 binds that record's hash.
After package verification the parent requires the same host flag and boot ID;
the child must still report no sandbox reboot marker. Both false/false and
true/true are valid. Either flag transition or a changed boot ID denies acceptance.
The reported `reboot_required` remains truthful, and
`reboot_required_introduced` is false. No reboot is initiated by these jobs.
Fresh planning exposes `host_reboot_required`. If it is true, the immutable
plan reason displayed for approval includes
“Host reboot already required; flag unchanged; no reboot in this job”.
Browser preparation and approval both verify the reason in the canonical
persisted plan payload. The new runtime uses the separate
`configured-reconcile-runtime/cases/` namespace and refuses to repeat the
historical configure-only operation.
The historical read-only verifier accepts only the exact saved 337 context,
signed action and old runtime bindings, original partial reference, 13 immutable
evidence files, APT 0/armed hook and original configure-only protocol. It checks
the complete installed snapshot twice, unchanged other packages, payload,
empty audit, idle locks, same boot and stable current host marker. It produces
a new failed classification with `effect_complete=true`,
`postconditions_passed=false` and `reboot_required=true`; it cannot manufacture
the missing before sample or a success record. See
[the helper contract](../tools/apt-transaction/HELPER-CONTRACT.md).
## Actually executed verification
All **81 Python tests passed without skips** in the offline, read-only Docker
development image as UID 1000. This included the real APT 2.6.2 protocol and
argument regressions with a fake dpkg collector, which changed neither host nor
copied package status. Targeted coverage includes both equal flag states,
both forbidden transitions, missing before evidence, changed boot, private
marker rejection, acceptance-record binding and strict historical failure
classification. These local tests supplement the native evidence below.
Python AST, browser JavaScript syntax and `git diff --check` also passed.
The qualified provider SHA256 is
`1dc0237dd8df9c9d22568bb1f6b047ede73bc2a5fca5c7fcedb7f24ab2a58b1e`;
the installed x86-64 helper is
`bbd2cf80a4e67b8a37ba3845c2bbed8b3c84d1eb01c1ac515a34577c8caa8ad0`
and agent is
`f40c7e2646e8122f5320c0909b0172ed8c749c218ec76b7ba3b8b397eb970671`.
The preserving native runtime upgrade completed. Its SSH response timed out;
an independent readback proved the completed report and all three installed
hashes, so it was not repeated. The probe and report certify runtime readiness,
not package acceptance. Normal RootSync then stored one append-only effect
reconciliation, preserved the original unknown bytes and closed the original
337 grant as failed. The actual Root ledger has **11 consumed, 0 unresolved,
0 faults**; central counts are **20 jobs, 20 outbox entries, 18 permits,
0 unresolved**. The earlier partial proof is unchanged and all forbidden
acceptance records remain absent.
Protected local witnesses, excluded from Git:
- `devstate/cloud-apt-recovery/configured-reconcile-runtime.json`
- `devstate/cloud-apt-recovery/configured-maintenance-completed-readback.json`
- `devstate/cloud-apt-recovery/capacity-recovery/configured-closure-3b7e66aba1b1495db1675eaab5117393.json`
- `devstate/cloud-apt-recovery/helper-configured-reconcile.json`
- `devstate/cloud-apt-recovery/agent-configured-reconcile.json`
At this checkpoint, the new provider's separately approved native `1.0`→`1.1`
upgrade and following no-change jobs remain outstanding. The historical
classification is not their acceptance and does not complete the Linux MVP.
## Subsequent actual native 1.1 acceptance
The next separate immutable plan passed its actual simulated dev/test WebUI
approval and native Cloud execution. Before approval0jobs/0outbox/0permits;
afterward exactly1/1/1 and0unresolved. Grant
`6b68c65b-8c0c-4cab-964d-e12a2d7f0626`, Root consumption12, succeeded with
`postconditions_passed=true`, independently verified installed fixture1.1,
payload SHA256`1523e9e982ff804c6d979e1bb2c6f3d7bae35307e8ce458bffc0d12a442b98da`
and native record`a88d6b10bbf6c36eaa8dee87eb1a381f2becc36297db8df2b4189791b9165506`.
The bound host flag was actually true before and after, introductionfalse,
truthful result`reboot_required=true`; no reboot was initiated. No OS/kernel
upgrade or change outside the harmless fixture is claimed.
Private verified checkpoint:
`devstate/cloud-apt-recovery/configured-reconcile-runtime/cases/apt_native_update/checkpoint.json`.
The two separate normal1.1-noop plans and remaining MVP gates are still open.
The first separate `ensure_package` plan subsequently ended **failed** with
`STATUS_FINGERPRINT`, grant`b7fbd250-c382-4d2f-950d-ae6001e4c48e`/Root13.
Its actual UI audit and counts1/1/1/0 are retained. It reused exactly the
pre-upgrade baseline and status fingerprint; the native update's verified
post-snapshot was different. The under90-second observation-age check alone
therefore did not establish post-update freshness. The immutable failed plan
is not a passed noop and is never reactivated. The minimal next preparation
must bind a newly stored passive installed1.1 observation to the accepted
post-snapshot; any new attempt needs its own pending0/0/0 and actual WebUI click.
No provider/helper change or replacement of the accepted1.1 record is needed.
Actual read-only preflight then proved the exact terminal Root13 outcome,
active handoff13 with no launch/hook/prechecks/child result, and retained native12
acceptance plus original31/337 proofs. Root13/0unresolved/0faults. Protected
report`configured-reconcile-runtime/noop-preflight-d140758efc3e421b91550d3dd52d5d7a.json`.
The fresh, separately WebUI-approved `ensure_package` then passed on the actual
Cloud VM: pending0/0/0 →1job/1outbox/1permit/0unresolved. Its newly stored passive
planning observation was after the accepted update read and matched the exact
verified installed1.1 post-snapshot. Result succeeded with independently met
postconditions, `changed=false`, `already_desired=true`,
`transaction_executed=false`, and `postcondition_already_met=true`.
The first failed plan is unchanged. Protected successful checkpoint:
`configured-reconcile-runtime/cases/no-change-retry/ensure_package/checkpoint.json`.
The separate `apply_updates` noop remains open.
Finally the separate `apply_updates` noop passed natively as well, grant
`c267f82b-328f-4d36-8d70-9aa81108646d`/Root15, same exact1.1 post-snapshot and
truthful reboot flag. Both noop results have complete postconditions and
explicitly no transaction. APT test policy was then closed, endpoint removed
and service stopped. This completes bounded fixture M4; the other MVP gates
remain open. Successful apply checkpoint is
`configured-reconcile-runtime/cases/apply_updates/checkpoint.json`.
The SDK fix is implemented and locally qualified: **29 pure harness guard
tests passed**. A no-change preparation now independently reads the completed,
verified native upgrade, then waits for a passive observation collected after
that read. It requires installed fixture `1.1` and the exact full baseline hash
from the native upgrade's postconditions. The earlier observation's under
90-second age is insufficient. A saved preparation witness records that time
barrier, observation time and baseline/fingerprint.
The explicit `--new-noop-after-definitive-no-effect` option additionally checks
the old immutable failed plan, its single authoritative grant and a read-only
Root-ledger/context proof that the `STATUS_FINGERPRINT` failure occurred before
launch or hook arming. It creates a separate
`cases/no-change-retry/ensure_package/` directory; the original checkpoint is
retained. The next actual browser flow still creates a new pending plan and
requires its own approval. These commands are **prepared, not executed by this
source qualification**:
```sh
python3 deploy/cloud_apt_recovery.py accept --case ensure_package --new-noop-after-definitive-no-effect
python3 deploy/cloud_apt_recovery.py finish --case ensure_package --new-noop-after-definitive-no-effect
```